i7 980X - QPI 2x or 1.5x

[inovice]

Hello,

I have just purchased a 2nd hand i7 980X to replace my i7 920. I have Asus P6T mobo with 6 GB of Memory and Zalman CNPS14X as heatsink. My 920 was OC-ed to 3.8 GHz which is largely enough for a noob like me

I have been doing a lot of reading lately. I understand Core i7 980X gulftown can utilise a lower uncore to memory ratio of 1.5x compared to Core i7 920 bloomfield's 2.0x ratio. Meaning you can theoretically use lower uncore multiplier and frequency.

But I have not been able to see any Intel documentation about this. Where does this assumption come from? Just curious to know, as Asus Bios mentioned it should be 2x

 

[lopri]

It's kind of a simple math so I don't think you need to dig for a document. When Nehalem debuted, with it came tripple-channel memory interface between the CPU and the memory. Its memory controller handled 92-bit (32 bit x 3) interface to L3. Lynnfield brought back dual-channel interface (48 bit x 2) between L3 and memory controller. The 2.0x ratio is still maintained internally but the 48 bit interface allows IMC/L3 to slow down to 1.5x memory rate. (48/32 = 1.5) Wider interface tends to let things to slow down while maintaining same transfer rates.

There is no assumption here, but a simple math.

Oh, and welcome to the AT Forums. It's always refreshing to meet a newcomer with intellectual curiosity.

 

[Brekyrself]

Here is a simple overclocking guide that shows the main differences with gulftown.

http://www.techreaction.net/2010/09/07/3-step-overclocking-guide-bloomfield-and-gulftown/2/

 

[inovice]

Hi lopri,

Thanks for your input and you are right :thumbsup: The information below is from another webpage, similar question and it was answered by an Intel Engineer back in 2010.

"Bloomfield i7 9xx offered three channels of DDR3-1066 support, Core i7-980X does that as well. But a number of the memory vendors have triple-channel kits rated for DDR3-2000 at 1.65V. When you initiate XMP profile for DDR3-2000, you are also increasing the uncore voltage from 1.2V to 1.6/1.7V, extremely aggressive for i7 980x (32 nm)

Starting with Lynnfield, we altered the ratio between the uncore and memory from 2:1 to 1.5:1. Thus enthusiasts who want to run DDR-2000 memory will not be forced to a 4 GHz uncore (requiring extreme voltages to achieved). Instead, you would need a 3 GHz uncore.

We dropped the ratio to allow faster memory speeds without needing high uncore voltage. Uncore frequency is simply a function of the silicon bin split (just like core frequency). If we only supported 2:1, that would limit the number of products in the pipeline. This is especially important for the mobile segment, with lower power/voltage for uncore on mobile platforms" (he is referring to Lynnfield/Clarksfield)

Thanks

 

[inovice]

Thanks Brekyrself! I read that couple of weeks ago before buying my Gulftown.

http://www.techreaction.net/2010/09/07/3-step-overclocking-guide-bloomfield-and-gulftown/2/

 

[james1701]

I read that article last night and decided to play with the uncore settings, and found that makes a large performance difference when lowering it. When I raised it, instead of lowering it, I started getting a massive boost in performance. From the articles I have read, they make the performance difference out to be negligible, and its not.

Using 3dmark 11 CPU test as a baseline, when I lowered uncore down to 1.5:1, I was getting 27.7 fps on that test. When I went back to a stock speed, and turned off HT, I was getting 26.8fps. When I raised it, above stock settings, I was getting 33.9fps on the CPU test with HT enabled.

After that, I ran 3dMark, and Crysis 1 benchmark, and had improvements all across the board. I got a 10fps bump in my Crysis benchmarks, and improvements all across the board in 3dMark, and Valley benchmark.

So, I was able to lower my overclock, get better performance, decrease voltage, and run a lot cooler. To be honest, I don't know if it was because I was bottlenecked with 2 7970's, or what, but I like the results.

Play with it if you have time.

 

[inovice]

Thanks james1701, I would start playing with my i7 980 over the weekend

I still do not understand why Intel move from a tri-channel to dual-channel memory controller for this cpu. I understand their concerns about power management for mobile devices. But why in the world would you do that to extreme processor series for desktop-the processor is aimed to satisfy enthusiasts who would most likely play around with voltage and memory multipliers (QPI/DRAM). The culprits are the memory makers for selling DDR3 2000 for X58 Chispet. As if 1.7V uncore memory on 45 nm Bloomfield cpu is perfectly fine but beware of the new 32 nm fab. Does not make any sense, although I do understand 32 nm processor are more sensitive to extra voltage.

I would perform some tests over the weekend

 

[james1701]

*Deleted*

 

[VirtualLarry]

It's kind of a simple math so I don't think you need to dig for a document. When Nehalem debuted, with it came tripple-channel memory interface between the CPU and the memory. Its memory controller handled 92-bit (32 bit x 3) interface to L3. Lynnfield brought back dual-channel interface (48 bit x 2) between L3 and memory controller. The 2.0x ratio is still maintained internally but the 48 bit interface allows IMC/L3 to slow down to 1.5x memory rate. (48/32 = 1.5) Wider interface tends to let things to slow down while maintaining same transfer rates.

There is no assumption here, but a simple math.

Oh, and welcome to the AT Forums. It's always refreshng to meet a newcomer with intellectual curiosity.

Why bring up lynnfield, isn't his chip a westmere?

 

[SocketF]

That 1.5x multiplier seems to be a bad decision, as james says uncore clock is important. In contrast to this, RAM speed is uninteresting, especially if you have 3 channels, not only 2.

I ran my cheap RAM at CL8 and DDR3-1200 only, but my uncore and thus the L3 cache runs @3.9 Ghz. Cache is the important factor, not the RAM. There is a very good reason why intel switched to run the L3@core clock with sandy bridge ...